Developmental differences in functional organization of multispectral networks

Abstract

Assessing brain connectivity during rest has become a widely used approach to identify changes in functional brain organization during development. Generally, previous works have demonstrated that brain activity shifts from more local to more distributed processing from childhood into adolescence. However, the majority of those works have been based on functional magnetic resonance imaging measures, whereas multispectral functional connectivity, as measured using magnetoencephalography (MEG), has been far less characterized. In our study, we examined spontaneous cortical activity during eyes-closed rest using MEG in 101 typically developing youth (9–15 years old; 51 females, 50 males). Multispectral MEG images were computed, and connectivity was estimated in the canonical delta, theta, alpha, beta, and gamma bands using the imaginary part of the phase coherence, which was computed between 200 brain regions defined by the Schaefer cortical atlas. Delta and alpha connectivity matrices formed more communities as a function of increasing age. Connectivity weights predominantly decreased with age in both frequency bands; delta-band differences largely implicated limbic cortical regions and alpha band differences in attention and cognitive networks. These results are consistent with previous work, indicating the functional organization of the brain becomes more segregated across development, and highlight spectral specificity across different canonical networks.

Introduction

The brain is an autonomously active system where the spontaneous activity of neural populations synchronize, composing a coherent functional architecture (Raichle 2010). Previous research, predominantly using functional magnetic resonance imaging (fMRI), has identified multiple networks of brain regions which exhibit synchronous activity (Calhoun et al. 2002; Fox and Raichle 2007; Doucet et al. 2011), and has shown that these networks take shape across development (Gu et al. 2015; Faghiri et al. 2018). During development, the organization of resting state networks appears to shift from local to more spatially distributed networks, with functional brain activity becoming more segregated from adolescence into adulthood (Fair et al. 2007, 2008; Kelly et al. 2009). Formation of resting networks through segregation is thought to be a critical feature of the development of brain function (Grayson and Fair 2017), and may underlie the profound changes in goal-directed cognition and social behavior that occurs during adolescence (Satterthwaite et al. 2015). At the same time, adolescence is also a particularly sensitive window of development given the dynamic changes in network topology that occur (Ernst et al. 2015; van Duijvenvoorde et al. 2016), with markedly increased risk for emergent psychopathology compared to other developmental time periods (Casey et al. 2008; Powers and Casey 2015). Further, altered development of resting state brain networks during this period have been associated with cognitive (McTeague et al. 2017) and behavioral dysfunction (Silveira et al. 2020; Zhang et al. 2021). Thus, characterizing how these resting state networks take shape across adolescence is critical toward understanding the underpinnings of neurocognitive and behavioral development and dysfunction.

While the majority of resting state studies in developing youth have measured connectivity using the blood oxygenation-level dependent (BOLD) signal, computing connectivity using the magnetoencephalography (MEG) time series provides a unique set of parameters that are ideal for developmental studies. The BOLD signal is a vascular measure related to blood oxygenation (Ogawa et al. 1990) that is only partially coupled with neural activity changes (Shmuel and Leopold 2008; Schölvinck et al. 2010), and may be influenced by neurovascular differences occurring independent of neural activity (Rack-Gomer et al. 2009; Das et al. 2021), including important age-related changes (D’Esposito et al. 2003; Burnett et al. 2011). Alternatively, MEG provides a direct measure of neural activity with high temporal and spatial precision, which is not affected by age-related changes to the neurovascular system. Further, MEG recording is silent and noninvasive, making it ideally suited for developing youth, who may be particularly sensitive to the noisy and restricted MRI environment. Beyond age considerations, whereas fMRI-BOLD indirectly infers neural activity vis-à-vis vascular change (Rack-Gomer et al. 2009; Das et al. 2021), MEG directly measures postsynaptic potentials within active neural populations (Hämäläinen et al. 1993; Wilson et al. 2016). Similarly, the temporal resolution of MEG is on the order of milliseconds, which permits the measurement of connectivity in frequencies across the brain’s intrinsic rhythms. In contrast, BOLD generally captures connectivity amid the infra-slow vascular changes [~0.1 Hz; (Drew et al. 2020; DeRamus et al. 2021)]. Thus, the connectivity measured by MEG and fMRI are built on fundamentally different biophysical properties. Taken together, MEG is an excellent complement to the collection of connectivity studies that have used fMRI-BOLD in youth.

Despite the clear advantages, work using MEG to study the development of multispectral brain connectivity is limited. The two studies that have used MEG to compute connectivity in source space both found age-related connectivity increases across multiple frequency bands, especially alpha (Schäfer et al. 2014; Hunt et al. 2019). These findings are suggested to be indicative of greater integration of distributed networks, which is consistent with some fMRI work (Fair et al. 2007). In addition, Hunt et al. (2019) used a graph theoretical measure of overall network connectivity to show that older adolescents show a stronger connectivity in the theta, alpha, and beta frequency bands. These studies, however, do not focus on adolescent ages, but rather compare children and adult groups. In addition, Schäfer et al. (2014) measured connectivity using amplitude envelope correlation, whereas Hunt et al. (2019) used phase coherence, both of which are measures that are susceptible to zero-lag phase correlations caused by volume conduction (Nolte et al. 2004). Notably, other studies conducted in (electroencephalography) sensor space have found results consistent with both of these MEG studies (Srinivasan 1999; Barry et al. 2004; Smit et al. 2012; Miskovic et al. 2015), while others have found contradictory results (van Baal et al. 2001; Micheloyannis et al. 2009; Boersma et al. 2011); but these results should be interpreted with caution given that computing connectivity at the scalp level can result in fully spurious connectivity estimates, depending on the specific metric used, whether the analysis was conducted at zero-lag, and other analytical elements (Brookes et al. 2011; Palva and Palva 2012). In sum, the sparse work using MEG to estimate network connectivity in adolescence supports some of the fMRI findings, but much remains to be discovered regarding the multispectral aspect, and studies focused on adolescent development in particular do not yet exist.

The goal of our study was to characterize the age-related differences in multispectral network connectivity across an adolescent age range (9–15 years). Whole-brain connectivity was quantified in each of the five canonical frequency bands (delta, theta, alpha, beta, and gamma) using imaginary phase-coherence (Nolte et al. 2004), which suppresses instantaneous phase synchrony caused by volume conduction and thus enhances the precision of connectivity estimates. The resulting connectivity matrices were first submitted to an analysis of mesoscale structure using a Louvain-like locally greedy algorithm, to determine if the network structure of connections across the brain become more segregated in older adolescents. This procedure has been widely used (Doron et al. 2012; Bassett et al. 2013; Braun et al. 2016; Doucet et al. 2017) to measure modularity by assigning regions into separate communities (i.e. modules) based on their mutual edge strengths. In line with the fMRI-BOLD connectivity literature which demonstrates the brain shifts from local to distributed processing with age, we predicted that network structure would be segregated into more communities in older adolescents. Frequency bands showing a relationship between age and community structure were further studied via an edge-wise regression with age to characterize how the age-related differences contribute to alterations in community structure.

Materials and methods

Participants

A total of 127 heathy youths (66 females, 61 males; 119 right-handed) were enrolled in the study (M_age = 11.78 years, SD = 1.60, range = 9.03–15.20 years). In addition to common demographics, socioeconomic status was assessed using the Barratt Simplified Measure of Social Status [BSMSS (Barratt 2006)], where the total score summarizes parental education, income, and occupational status. Data from this sample was included in a recently published study examining developmental changes in resting spontaneous cortical power (Ott et al. 2021). A parent or legal guardian provided informed consent and reported all demographic information. The child participants provided assent before participating in the study. The study was fully approved by our Institutional Review Board, and all protocols were conducted in accordance with the Declaration of Helsinki.

Exclusionary criteria included the inability to complete the resting state MEG scan, any medical illness or medication affecting CNS function, neurological, or psychiatric disorders, history of head trauma, current substance abuse, and the standard exclusion criteria related to MEG and MRI acquisition (e.g. dental braces, metallic implants, battery operated implants, and/or any type of ferromagnetic implanted material).

MEG acquisition

The MEG signals were recorded from a 306-sensor Elekta/MEGIN MEG system (Helsinki, Finland), equipped with 204 planar gradiometers and 102 magnetometers, which sampled the neuromagnetic responses continuously at 1 kHz with an acquisition bandwidth of 0.1–330 Hz. Recordings took place inside a one-layer magnetically shielded room with active shielding engaged for environmental noise compensation. Participants were seated in a custom-made nonmagnetic chair, with their heads positioned within the sensor array.

Participants were instructed to rest with their eyes closed throughout a 6-min block. Participants were monitored throughout MEG data acquisition via live audio–video feeds inside the shielded room. Structural T1 weighted images were acquired on a Siemens 3 T Skyra scanner with a 32-channel head coil using a MPRAGE sequence (TR = 2,400 ms; TE = 1.94 ms; flip angle = 8°; FOV = 256 mm; slice thickness = 1 mm; base resolution = 256; 192 slices; voxel size = 1 × 1 × 1 mm).

Structural MRI processing and MEG–MRI co-registration

Participants’ high-resolution T1-weighted structural MRI data were segmented using a standard voxel-based morphometry pipeline in the computational anatomy toolbox [CAT12 v12.7; (Gaser et al. 2016)] within SPM12. Segmented T1 images underwent noise reduction using a spatially adaptive nonlocal means denoising filter (Manjón et al. 2010) and a classical Markov Random Field approach (Rajapakse et al. 1997). An affine registration and a local intensity transformation were then applied to the bias corrected images. These preprocessed images were segmented based on an adaptive maximum a posteriori technique (Ashburner and Friston 2005) and a partial volume estimation with a simplified mixed model of a maximum of two tissue types. Lastly, the segmented images were normalized to MNI template space and imported into Brainstorm for co-registration.

Prior to MEG acquisition, four coils were attached to the participants’ heads and localized with the three fiducial points and scalp surface using a 3-D digitizer (Fastrak 3SF0002, Polhemus Navigator Sciences, Colchester, VT, USA). After the participant was positioned for MEG recording, an electrical current with a unique frequency label (e.g. 322 Hz) was fed into each of the coils, which induced a measurable magnetic field and allowed each coil to be uniquely localized relative to the sensors throughout the recording session. Here, because the coil locations were also known in head coordinates, all MEG measurements could be transformed into a common coordinate system. This coordinate system was then used to co-register each participant’s MEG data to their structural MRI, prior to source space analyses using Brainstorm (see MEG source imaging and frequency power maps).

MEG data preprocessing

Each MEG dataset was individually corrected for head motion and subjected to noise reduction using the signal space separation method with a temporal extension [tSSS; MaxFilter v 2.2; correlation limit: 0.950; correlation window duration: 6 s; (Taulu and Simola 2006)]. The preprocessing of MEG data was conducted in Brainstorm (Tadel et al. 2019) and was modeled after that of previous analyses of resting state data (Niso et al. 2019; Ott et al. 2021; Petro et al. 2022). A high pass filter of 0.3 Hz and notch filters at 60 Hz and its harmonics were applied. Cardiac artifacts were identified in the raw MEG data and removed using an adaptive signal space projection approach, which was subsequently accounted for during source reconstruction (Uusitalo and Ilmoniemi 1997; Ille et al. 2002). Following artifact removal, data were divided into 4-s epochs, which were examined for quality on a per-person basis. Specifically, individual thresholds for rejection of epochs were calculated based on both the amplitude and gradient distribution of the signal, given the amplitude of the MEG signal is affected exponentially by the distance between the brain and MEG sensor array. Following artifact rejection, the average number of epochs retained was 55.82 (SD = 6.89, min = 40, max = 71) per participant, and did not vary as a function of age or sex (ps > 0.05).

MEG source imaging and frequency power maps

Source modeling followed the analysis pipeline outlined in Wiesman et al. (2021). Briefly, the forward model was computed using an overlapping spheres head model (Huang et al. 1999), unconstrained to the cortical surface. A linearly constrained minimum variance beamformer, implemented in Brainstorm, was then used to spatially filter the data epochs based on both the data covariance, computed from the resting state recording, and the noise covariance, computed from the recordings of the empty room. The vertex-wise source time series data were then quantified within 200 parcellations using the Shaefer atlas (Schaefer et al. 2018), and a principle component analysis (PCA) was used to extract the average signal among all vertices within each parcellation. This atlas was selected because its regions have been separated into seven canonical networks by previous connectivity studies (Yeo et al. 2011): cognitive control, dorsal attention, default mode, limbic, somato-motor, salience/ventral attention, visual.

MEG oscillatory connectivity

Using the parcellation-wise time series, functional connectivity was computed using the imaginary component of the coherency [i.e. imaginary phase coherence; (Nolte et al. 2004)], computed in the Brainstorm toolbox (Tadel et al. 2019). This method of calculating functional connectivity suppresses relationships with zero millisecond phase lag, and thus is immune to artificial connectivity caused by linear mixing of signals among spatially proximate sources. Separately for each pair of the 200 parcellation time series (i.e. between each region in the Schaefer atlas), the imaginary phase coherence was calculated for each frequency f as:

Where S_xy is the cross power spectral density and S_xx and S_yy are the power spectral densities of either time series x or y, respectively. The power and cross power spectral density were estimated using Welch’s method with a Hanning window of 1 s duration with 50% overlap, parameters known to prevent the impact of signal nonstationarity on the connectivity estimates (Bortel and Sovka 2007). Using this method, connectivity was measured on each of the artifact free epochs, which was then averaged across epochs for each person and frequency band separately. This method resulted in a 200 × 200 matrix of connectivity estimates, per person, for each of the five canonical frequency bands: delta (2–4 Hz), theta (5–7 Hz), alpha (8–12 Hz), beta (15–29 Hz), gamma (30–59 Hz).

Mesoscale community structure and age

The community structure of each participant’s 200 × 200 connectivity matrix was calculated using a Louvain-like locally greedy algorithm implemented in MATLAB (Jeub et al. 2011), which is a data-driven approach that assigns each node to a constituent module (i.e. a community) based on the density of interconnectivity with other nodes. This procedure was implemented similar to that in several previous studies (Doron et al. 2012; Bassett et al. 2013; Braun et al. 2016; Doucet et al. 2017) and identifies the optimal partition of nodes into modules m by searching for a partition that maximizes the following modularity quality function:

where m_i is the module containing node i, w is the sum of all weights, and = 1 if node i and node j are in the same module or otherwise = 0. Here, a resolution parameter ɣ of 1 was used, in line with the common choice across the literature (e.g. Bassett et al. 2013; Tang et al. 2017). Because this procedure considered nodes in a pseudorandom order, different results can occur from run to run. Therefore, this procedure was conducted 100 times per participant and a single consensus map was calculated from these 100 community assignment maps (Lancichinetti and Fortunato 2012). Here, the z-score of the Rand coefficient was used to determine the most similar partition among all 100 of the iterative maps, for each node separately. This single consensus map was used for subsequent analysis.

For each participant, the size of the largest community and total number of communities were used as measures of the degree to which the functional connections could be partitioned into separate communities; these measures, separately for each frequency band, were compared with age. In other words, these analyses tested if the functional architecture of spontaneous cortical activity differed as a function of age. These comparisons with age were accomplished by submitting community size and number of total communities to separate regression models as the outcome variable, with age as the predictor and sex as a covariate of no interest.

Edge-wise connectivity and age

In addition to describing the mesoscale community structure of functional connectivity, the relationship between age and each individual edge in the 200 × 200 connectivity matrix was tested. This was accomplished using an edge-wise regression where each edge weight served as the outcome variable and age as the predictor variable; in addition, sex was included as a covariate of no interest. This process was conducted for each frequency band showing a relationship between age and community structure (see MEG oscillatory connectivity).

To account for multiple comparisons, significance thresholds for each edge were determined via permutation testing. Here, the edge weights and predictors were randomly shuffled across observations before computing a permuted statistic. This process was repeated 10,000 times, separately at each edge. Thus, a unique distribution of regression coefficients was generated at each edge, to which the age-related regression coefficients were compared. The regression coefficient at any given edge was determined to be significant if exceeding the 0.25th percentile of either end of its permuted distribution (i.e. passed a P < 0.005 threshold).

Results

Descriptive statistics

Of the 127 enrolled participants, nine failed to complete the MRI and 10 failed to complete the MEG resting state task. In addition, the connectivity data from seven participants were excluded as their overall connectivity (across all edges) exceeded 2.5 standard deviations of the mean in at least one of the five frequency bands. Thus, the current study included the analysis of full data from 101 participants (M_age = 11.88, SD = 1.64, range = 9.03–15.20 years). Demographic characteristics are detailed in Table 1. Socioeconomic status tended to the higher range [BSMSS = 46.36(12.43)] of possible scores (8–66).

Table 1.

Demographic characteristics of the final sample.

	Male	Female	P-Value
Age Range (years)	9.03–14.85	9.34–15.20	–
Mean Age (years)	12.02	11.76	0.42
Race (White/Black or African American/Other/Unknown)	45/1/1/3	38/4/6/3	0.11
Ethnicity (Not Hispanic or Latino/Hispanic or Latino/Unknown)	45/5/0	46/4/1	0.57
Handedness (R/L/both)	46/3/1	49/2/0	0.53
Socioeconomic Status (BSMSS; Barratt 2006)	48.01	46.72	0.61

Note: Differences in mean age between males and females were assessed using an independent samples t-test; differences in race, ethnicity, and handedness were assessed using chi-square tests.

Louvain-like model

Using the Louvain-like locally greedy algorithm, the edge weights were used to partition nodes to communities, separately for each participant. The number of communities was positively related to age only in the delta (t₉₈ = 2.03, P < 0.05) and alpha (t₉₈ = 2.18, P < 0.05) bands. Similarly, the size of the largest community was negatively related to age for only the delta (t₉₈ = −2.44, P < 0.05) and alpha (t₉₈ = −2.68, P < 0.01; Fig. 1) frequency bands.

Fig. 1.

Developmental differences in network community. Age (x-axes) is plotted against the size of the largest community (left column) and the total number of communities (right column), separately for connectivity in the delta (top) and alpha (bottom) frequency bands. The least squares line is plotted in black, and the shaded region illustrates the 95% confidence interval. For both delta and alpha, the network structure was divided into more numerous and smaller communities in older adolescents.

The modularity coefficient computed from the Louvain-like algorithms was also submitted to a regression with age, but it was not related to age for either delta or alpha connectivity (ps > 0.05).

Edge-wise regression

To identify the edges contributing to age-related differences in the alpha and delta connectivity networks, the connectivity matrices were submitted to an edge-wise regression, where edge weight was the outcome variable with age as the predictor. Significance of each edge’s regression was then determined by a permuted threshold. To describe the contribution of the networks defined by the Schaefer atlas, the standardized beta weights were averaged for the edges of each network separately. Thus, the averaged standardized beta weights represent the strength of the effect of age in each canonical network.

In the delta band, a total of 29 edges were negatively related to age, and no edges were positively related to age (see supplemental Table S1 for the number of edges in each network). The standardized betas for the negative effect of age were largest for edges with nodes in the limbic network, specifically temporal pole regions. In particular, the standardized betas were strongest among the limbic-to-limbic edges. Figure 2 shows each brain region with at least one edge where delta connectivity was significantly related to age (panel A), and the degree to which each network was implicated in the age effect (panels B and C).

Fig. 2.

Illustration of region and network level relationship between delta connectivity and age. (A) Illustration of brain regions exhibiting negative relationships with age. Edge weights related to decreasing age, after permuted significance thresholding, where stronger negative relationships are illustrated with thicker lines and more magenta (compared to yellow) color. Nodes are illustrated as spheres, where larger sizes indicate a greater number of significant edges; nodes are colored by network according to the color labels indicated in panel B. (B) Illustration of the number of edges negatively related to age. The chord diagram shows the network-wise distribution of significant age-related edges among the networks defined in the Schaefer atlas: Cognitive control (Cont), dorsal attention (DorsAttn), default mode (Default), limbic (Limbic), somato-motor (SomatoMotor), salience/ventral attention (VenAttn), visual (Vis). The networks are ordered clockwise by the total number of significant edges. Size of each chord reflects the number of connections between (or within) networks. Note that all connections are bidirectional; chord colors are differentiated for illustrative purposes and are colored according to the network with the overall greatest number of significant edges. Colored caps are included at the ends of each chord to indicate the opposite network shared by each chord. (C) Illustration of the strength of the negative age relationship in each network. Heatmap colors indicate the average standard beta weights of all network-wise edges. Adjacent bar graphs reflect the average standardized Beta weights of edges implicated in each network (i.e. the average of standardized beta weights in each row).

In the alpha band, a total of 109 edges were negatively related to age and one edge was positively related to age (between two nodes in the default network; see supplemental Tables S2 and S3 for the number of edges in each network). The standardized betas for the negative effect of age were largest for edges with nodes in the ventral attention network, and among ventral attention to visual edges in particular. Figure 3 illustrates each brain region with at least one edge where alpha connectivity was significantly related to age (panel A), and the degree to which each network was implicated in the age effect (panels B and C).

Fig. 3.

Illustration of region and network level relationship between alpha connectivity and age. (A) Illustration of brain regions exhibiting negative relationships with age. Edge weights related to decreasing age, after permuted significance thresholding, are shown with stronger negative relationships indicated by thicker lines and more magenta (compared to yellow) color. Nodes are illustrated as spheres, where larger sizes indicate a greater number of significant edges; nodes are colored by network according to the color labels indicated in panel B. (B) Illustration of the number of edges negatively related to age. Chord diagram showing the network-wise distribution of significant age-related edges among the networks defined in the Schaefer atlas: cognitive control (Cont), dorsal attention (DorsAttn), default mode (Default), limbic (Limbic), somato-motor (SomatoMotor), salience/ventral attention (VenAttn), visual (Vis). The networks are ordered clockwise by the total number of significant edges. Size of each chord reflects the number of connections between (or within) networks. Note that all connections are bidirectional; chord colors are differentiated for illustrative purposes and are colored according to the network with the overall greatest number of significant edges. Colored caps are included at the ends of each chord to indicate the opposite network shared by each chord. (C) Illustration of the strength of the negative age relationship in each network. Heatmap colors indicate the average standard beta weights of all network-wise edges. Adjacent bar graphs reflect the average standardized Beta weights of edges implicated in each network (i.e. the average of standardized Beta weights in each row).

Discussion

The current study examined age-related differences in whole-brain, multispectral connectivity within a sample of over 100 children and adolescents (9–15 years) who underwent a resting state MEG session. The MEG signals were projected onto 200 parcellated regions using the Shaefer atlas and examined for network architecture. Our results suggested smaller and more numerous network communities in older adolescents in the delta and alpha bands, indicative of a functional architecture that is more segregated. These age-related differences in network architecture were accompanied by predominantly weaker connectivity in older adolescence and involved different brain regions between the delta and alpha bands. In the delta range, this age effect largely implicated limbic cortical regions, whereas in the alpha band these decreases involved the cognitive control and attention networks.

The current findings are generally consistent with the notion that brain regions increasingly segregate during adolescence (Gu et al. 2015; He et al. 2019). The Louvain-like community detection algorithm is a data-driven approach to assign nodes into separate communities based on the density of mutual connections, where a greater number of communities is indicative of a greater number of unique sets of interconnected nodes (Jeub et al. 2011). Thus, the increased number of communities in older relative to younger adolescents suggests that the architecture of functional brain networks becomes increasingly segregated across adolescence. These results are in line with previous works demonstrating that segregation increases as a function of development (He et al. 2019; Wang et al. 2020) and supports cognitive development (Baum et al. 2017). In addition, the edge-wise regression found numerous decreases in connectivity strength in older adolescents. Together, these results suggest that the increased segregation in functional architecture is driven by decreases in connectivity strength.

While older adolescents exhibited increased segregation in both delta and alpha ranges, age-related differences in modularity were not observed. In graph theory, modularity is defined by strong, dense local connections but sparse connections between communities (Newman and Girvan 2004). The edge-wise analysis of the current data indicated numerous negative relationships with age, but positive relationships with age were sparse. Thus, the lack of modularity differences across age can be seen as driven by the lack of strengthening connections with age. One possibility is that strengthening connections with age may occur at more local connections (i.e. at the vertex-wise level within parcellated regions), which are obscured by our current atlas-based parcellation approach as it collapses across such vertices. We selected the 200-region atlas approach over vertex-wise connectivity in the current study to leverage the predefined, well-known network labels and ease computational demands, thus enabling replicability among other research groups. Future work should probe the degree to which vertex-wise connectivity may indicate local connectivity changes.

While increased segregation is in line with previous fMRI work, the lack of age-related differences in modularity is at odds with some fMRI literature. Many earlier fMRI studies found that, inconsistent with our findings, local and long-range connections decreased and increased in older adolescents, respectively (Fair et al. 2009; Dosenbach et al. 2010). However, subsequent studies demonstrated that this age-related effect in fMRI is confounded with motion artifacts that are more prevalent in younger ages (Satterthwaite et al. 2012; Fair et al. 2013; Power et al. 2015), and that the pattern of connectivity differences in older youth was diminished when appropriately controlling for motion. The risk of motion contamination is reduced in the current MEG data given motion correction procedures applied to the raw time series data and the removal of artifact contaminated epochs. Moreover, the connectivity measurements involved a phase-coherence measure that excludes zero-lag coupling, which may be caused by artifactual nonbrain (e.g. muscular) sources (Miljevic et al. 2022). Notwithstanding the methodological differences in artifact handling, the biophysical properties that produce the MEG and fMRI signals are inherently different: MEG signal is likely generated from the postsynaptic potentials in neural populations, whereas fMRI-BOLD relies on indirect vascular changes (Rack-Gomer et al. 2009; Das et al. 2021). Relatedly, fMRI measures connectivity at a dramatically slower scale (i.e. at ultra-low frequencies), and thus may capture fundamentally distinct connectivity. Lastly, the precise definitions of local and long-range connectivity are not consistent across studies (see Picci et al., 2016, for a review) making their comparison difficult across studies. Given these differences, the discrepancy between the current results and earlier fMRI work is unsurprising.

Age-related connectivity differences implicated different canonical networks between delta and alpha. In the delta band, age-related differences primarily involved limbic cortical regions in the temporal pole, which are intrinsically coupled with numerous aspects of complex phenomena including social, affective, and memory processes (e.g. mentalizing, prosociality; (Do et al. 2019; Fehlbaum et al. 2021; Sahi and Eisenberger 2021; van Buuren et al. 2021). Indeed, these regions undergo considerable structural and functional changes through adolescence into adulthood in order to accommodate the increasingly complex social environments individuals must navigate (Andrews et al. 2021; Sydnor et al. 2021). The current results suggest that one aspect of this development involves decreasing connectivity. In the alpha band, age-related differences were relatively more widespread, but the effect was strongest for edges between ventral attention and visual regions. This result is unsurprising given the long held relationship between alpha and visual attention (Sauseng et al. 2005; Klimesch et al. 2007), and the known development of attention processes that occur during this time (Rueda et al. 2004; Taylor et al. 2021). Taken together, age-related decreases in connectivity like those observed here may underlie the developmental changes involved in both visual attention and more complex functions like emotion and memory processes.

Before closing, it is important to note the limitations of the current study, which may be addressable in future work. First, a cross-sectional design was used in the current study and future work using longitudinal designs may enable stronger conclusions. This is especially true given the high reliability of resting state MEG maps over at least 3-year time periods (Lew et al. 2021). Similarly, future work should measure pubertal stage [e.g. the Sexual Maturation Scale (Morris and Udry 1980) or Pubertal Development Scale (Petersen et al. 1988)] and/or directly measure pubertal hormones such as estradiol and testosterone (Fung et al. 2020, 2022a, b; Penhale et al. 2022), rather than relying on chronological age to help link connectivity changes to the neurobiological events that occur during puberty. Further, the somewhat restricted demographics of the current sample may limit the generalizability of the findings. Despite the strengths of using resting state designs, the extent to which findings can be tied to specific cognitive or emotional functions are limited, and future work should test if the current results generalize to task-induced connectivity using tasks that more directly probe cognitive and emotional processing. Notably, future work should carefully consider the use of eyes-open and -closed rest conditions, given the known differences in brain activity (Agcaoglu et al. 2020; Petro et al. 2022), and consider the potential impact of developmental cortical thickness changes on MEG measures of neural activity and connectivity. Previous studies investigating the relationship between cortical thickness and oscillatory activity have found mixed results, with some reporting a relationship (Gaetz et al. 2011; Wilson et al. 2015; van Pelt et al. 2018; Proskovec et al. 2020; Casagrande et al. 2021; Trevarrow et al. 2022; Springer et al. 2023) and others not (Edden et al. 2009; Schwarzkopf et al. 2012; Robson et al. 2015). In addition, a recent study from our lab found that cortical thickness was not a better predictor of motor beta power than age alone (Rempe et al. 2022), suggesting that oscillatory activity reflects unique age-related changes.

To summarize, the current study found age-related differences in resting connectivity within the delta and alpha band, such that older adolescents showed more segregated functional architecture. This large-scale difference in functional architecture was accompanied by numerous age-related decreases in connectivity, suggesting that the increased segregation is driven by the weakening of specific connections in older adolescence. In the delta band, these weakened connections predominantly implicated the limbic cortical areas, whereas weaker connections in the alpha band involved many frontoparietal cortical regions. These findings are generally consistent with the notion that brain networks become more functionally segregated during the formative years of adolescence. In addition, the current results point to intriguing connectivity differences between MEG and fMRI-BOLD methods.

Funding

This work was supported by the National Science Foundation (#1539067 to TWW, YPW, JMS, and VDC and #2112455 to VDC), the National Institutes of Health (R01MH121101, R01MH116782, R01MH118013, and P20-GM144641 to TWW; R01EB020407 and R01MH118695 to VDC; R56 MH124925 to YPW). Funding agencies had no part in the study design or the writing of this report.

Conflict of interest statement

All authors report no biomedical financial interests or potential conflicts of interest.

Data and code availability statement

The data used in this article will be made publicly available through the COINS framework at the completion of the study (https://coins.trendscenter.org/). Data processing pipelines followed previous studies (Niso et al. 2019) using a combination of Brainstorm (Tadel et al. 2019), which is documented and freely available for download online under the GNU general public license (http://neuroimage.usc.edu/brainstorm), and CAT12 (Gaser et al. 2016) toolboxes.

CRediT author statement

Nathan Petro (Conceptualization, Data curation, Formal analysis, Software, Visualization, Writing—original draft, Writing—review & editing), Giorgia Picci (Conceptualization, Writing—review & editing), Christine Embury (Formal analysis, Methodology, Software), Lauren R. Ott (Software, Writing—review & editing), Samantha H. Penhale (Software), Maggie P. Rempe (Methodology, Software, Writing—review & editing), Hallie J. Johnson (Data curation, Investigation, Project administration), Madelyn P. Willett (Data curation, Investigation, Project administration), Yu-Ping Wang (Conceptualization, Funding acquisition, Writing—review & editing), Julia Stephen (Conceptualization, Funding acquisition, Writing—review & editing), Vince Calhoun (Conceptualization, Funding acquisition, Writing—review & editing), Gaelle Doucet (Conceptualization, Formal analysis, Methodology, Supervision, Writing—review & editing), and Tony W. Wilson (Conceptualization, Funding acquisition, Project administration, Resources, Supervision, Writing—review & editing)